Causal Discovery and Forecasting in Nonstationary Environments with State-Space Models.

Area of Science:

• Economics
• Neuroscience
• Time Series Analysis

Background:

• Nonstationary time series present challenges for causal discovery and forecasting.
• Identifying causal relationships and predicting future values are crucial in many scientific fields.

Purpose of the Study:

• To investigate causal discovery and forecasting methods for nonstationary time series.
• To demonstrate how nonstationarity can facilitate causal structure identification and improve forecasting accuracy.

Main Methods:

• Utilized nonlinear state-space models to represent nonstationary processes.
• Allowed for time-varying causal strengths and noise variances within the models.
• Treated forecasting as a Bayesian inference problem within the learned causal model.

Main Results:

• Nonstationarity was shown to be beneficial for identifying causal structures.
• The proposed state-space models rendered causal structure and model parameters identifiable.
• Forecasting methods effectively exploited time-varying data properties and adapted to new observations.

Conclusions:

• The developed methods offer a principled approach to causal discovery and forecasting for nonstationary time series.
• Experimental results on synthetic and real-world data confirm the efficacy of the proposed techniques.
• This work provides valuable tools for analyzing complex dynamic systems in economics and neuroscience.